Learning Physically Consistent Lagrangian Control Models Without Acceleration Measurements